From c2e39bd488cf9041c4d9ae20cd2f464ece7efc61 Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Fri, 31 Mar 2023 17:23:25 +0800 Subject: [PATCH] Bug fix --- egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py index 2c215f0f1..18b58da92 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py @@ -301,7 +301,7 @@ class Zipformer2(EncoderInterface): max_downsampling_factor = max(downsampling_factor) - num_frames_max = (num_frames0 + max_downsampling_factor - 1) + num_frames_max = (num_frames0 + max_downsampling_factor - 1) // max_downsampling_factor # we divide the dropped-out feature dimensions into two equal groups; # the first group is dropped out with probability 0.1, the second @@ -325,7 +325,7 @@ class Zipformer2(EncoderInterface): feature_masks = [] for i in range(num_encoders): - ds = downsampling_factor[i] + ds = self.downsampling_factor[i] upsample_factor = (max_downsampling_factor // ds) frame_mask = (frame_mask_max.unsqueeze(1).expand(num_frames_max, upsample_factor,